9004 Problems of information security. Computer systems Проблемы информационной безопасности. Компьютерные системы 2071-8217 4 10.48612/jisp/mbvv-n1u7-z7be From exploitation to protection: a deep dive into adversarial attacks on LLMS От эксплуатации к защите: анализ атак на большие языковые модели 0009-0005-6662-5606 Velichko Ivan wwr0ngn4m3@gmail.com 0000-0002-0924-6221 Bezzateev Sergey sergey.bezzateev@gmail.com Saint Petersburg State University of Aerospace Instrumentation 25 03 2025 1 43 58

Modern large language models possess impressive capabilities but remain vulnerable to various attacks that can manipulate their responses, lead to leakage of confidential data, or bypass restrictions. This paper focuses on the analysis of prompt injection attacks, which allow bypassing model constraints, extracting hidden data, or forcing the model to follow malicious instructions.

 Large language models artificial intelligence adversarial attacks defense methods model output manipulation